To benefit the blind by using advanced deep learning techniques, we establish a new outdoor object detection dataset, BLIND. Different from PASCAL VOC, the scales of objects in our dataset are quite various because of different distances between objects and the camera. The characteristic of the BLIND dataset requires a high ability of scale invariance for the object detector, which classical SSD isn’t adequate. We propose a novel object detector named Feature Pyramid SSD (FPSSD) focusing on BLIND, applying feature fusion strategies to classical SSD. FPSSD achieves 75.4% mean Average Precision (mAP) on BLIND, surpassing classical SSD by 1.7%. Extensive experimental results and analyses demonstrate the necessity to establish the BLIND dataset and validate the effectiveness of the proposed FPSSD object detection algorithm for the blind people.